An expert interview on research into AI-driven drug optimization with Dr. Nils Weskamp, Associate Director of Computational Chemistry at Boehringer Ingelheim, and Dr. Thomas WollmannCTO at Merantix Momentum. Moderated by Dr. Gillian HertleinStrategic Project Manager at Merantix Momentum.

Gillian: Thank you to both of you for joining us today. First, could you briefly introduce yourself and provide a brief overview of the main factors determining whether a molecule is fit for further development and how these factors are traditionally evaluated and optimized?

Nils: Thank you, Gillian. I work as a team lead at Boehringer Ingelheim in the computational chemistry group in Biberach, Germany. We apply a broad range of machine learning techniques and modern simulation techniques to help our colleagues design great molecules.
In drug compound optimization, the goal is to create a molecule that is not only safe and effective but also convenient to use for patients. Balancing various molecular properties, such as potency, selectivity, solubility, metabolic stability and membrane permeability, is crucial. Traditionally, this process involves an iterative cycle of designing compounds, synthesizing samples, conducting experiments, and analyzing data to inform the next learning cycle.

Gillian: Moving on to the buzzwords of the day, ML and AI, Nils, how have these technologies changed drug compound enhancement compared to older methods?

Nils: ML methods excel in handling large datasets, allowing us to scale our optimization efforts. The speed at which ML models can make predictions enables faster iterations, transforming our working style. They not only scale traditional approaches but also reveal new principles by integrating disparate data types and discerning complex patterns beyond human capabilities.

Gillian: It's fascinating to see how compound optimization has evolved over time. Thomas, I understand you're working on a platform to streamline complex computational tasks. So to say, it enables scientists without a data science background to leverage the advantages of ML. Could you share more about the Galaxy Bioinformatics platform and your AI project in bioinformatics?

Thomas: The Galaxy Bioinformatics platform aims to provide high-performance computing resources to domain experts, fostering collaboration in this field. In my pet project, I'm developing an agent for bioinformatics to equip the Galaxy platform with an intuitive interface. This allows researchers to interact easily, query data, and gain insights without extensive computational expertise. Integrating AI agents into platforms like Galaxy could significantly enhance bioinformatic workflows like lead optimization.

Gillian: Nils, could you break down the step-by-step process of enhancing drug compounds using ML, from gathering and preparing data to choosing and fine-tuning models?

Nils: Most pharmaceutical companies have set up processes to ensure experimental results feed into a central database. Once we build a new model, we discuss dataset composition, class boundaries, and model application. The process involves automation, including structure normalization, featurization, and applying various ML methods. Time-based validation is crucial due to input space drift.

Gillian: Which ML algorithms or techniques are promising in compound optimization, and how are they improving efficiency?

Nils: Currently, classical tree-based methods like random forests stay prevalent. They are updated frequently, require minimal tuning, and deliver satisfactory results. There's a shift toward deep learning, especially in multitask learning, as it handles correlated parameters effectively.

Gillian: Thomas, could you elaborate on specific ML models impacting drug optimization and recent trends in this field?

Thomas: Generative models, graph neural networks, and transformer-based models are making waves. Generative models can be used to design novel compounds and explore chemical spaces humans might overlook. Graph neural networks can capture the structure of molecules directly. Transformer-based models can use muli-omics and clinical data while leveraging recent advancements in NLP. This so-called multi-modal AI can provide a more comprehensive understanding of disease mechanisms and drug responses.

Gillian: Moving to challenges, Nils, what are the main hurdles in integrating ML into compound optimization workflows, and how can these be overcome?

Nils: Dealing with experimental errors, noise, and biases in training data poses challenges. Models may not achieve 99% accuracy, necessitating consideration of applicability domains and uncertainty quantification. Explainable AI is crucial because our colleagues want to understand model patterns and compare them to their findings.

Gillian: Thomas, explainable AI is indeed crucial. What is explainable AI, and what approaches are currently explored in the field of explainable AI?

Thomas: Explainable AI refers to providing understandable and interpretable explanations for the decisions or predictions of artificial intelligence systems. It aims to create trust by demystifying complex models and making the decision-making process more transparent and understandable. Approaches include prototypes, feature importance, local explanations, counterfactual explanations, and interactive methods like those recently seen in LLMs.

Gillian: Nils, could you share examples of medical fields or diseases where ML-based compound enhancement has shown substantial success? And vice versa, where traditional methods were more effective than ML in compound optimization?

Nils: ML applications in drug discovery have been ongoing for years. While ML is excellent at interpolation, it shines in target classes with ample training data. Projects utilizing AI/ML have accelerated drug discovery, especially when extensive data analysis is needed.
ML faces challenges in unexplored areas without ample data, and traditional approaches outperform ML. Human experts excel in such scenarios, leveraging background knowledge and mechanical modeling. While ML is powerful, there are still unknowns requiring hypotheses and experiments.

Gillian: Based on your experiences, Nils, what advice would you give others regarding the learning curve of using ML for drug optimization?

Nils: Technology is just one aspect; the real challenge is deploying and making ML models accessible. Overcoming digital change management hurdles is crucial. Address concerns around AI/ML, promote lighthouse cases, educate on regulations, and emphasize the importance of data governance.

Thomas: I agree with Nils. Education on regulations like GDPR and the EU AI Act and the importance of data governance are especially crucial for AI implementation. These factors help build trust and navigate challenges and are also part of our Merantix AI School.

Gillian: Nils and Thomas, looking towards the future, which trends or technologies in ML and computational biology do you think will impact compound enhancement?

Nils: Integrating AI/ML with classical mechanistic modeling and simulation techniques holds potential. Additionally, despite challenges, generative AI could revolutionize workflows, enabling chemists to direct intelligent systems with their knowledge and feedback.

Thomas: AI will become even more present, acting as a companion to individuals in R&D. Collaborative synergy between human expertise and AI will redefine roles, shifting from direct engagement with raw data to interpreting and making strategic decisions based on ML-generated insights.

Gillian: Thank you for your valuable insights.